A Survey on Multi-Tenant QoS Support for SSD

Reading and summarizing research papers on multi-tenant QoS support for SSD.

Software Solutions

• D Shue, et al. Performance Isolation and Fairness for Multi-Tenant Cloud Storage, OSDI 2012
  • Contribution
    • Proposing the first system to provide per-tenant fair resource sharing support for cloud storage service.
  • Approach
    • Data partitions should be placed with respect to demand and node capacity constraints.
    • Weights for local access should give tenants throughput where they need it most.
    • Data replicas should be selected in a weight-sensitive manner during accesses.
    • Request queuing should enforce dominant resource fairness (DRF), e.g. from bytes in, bytes out and requests per second.
  • Evaluation
    • Evaluating MMR (min-max ratio) of the dominant resource across all tenants, 0 is worst and 1 is best.
• S Park, et al. FIOS: A Fair, Efficient Flash I/O Scheduler, FAST 2012
  • Contribution
    • FIOS, a new Flash I/O scheduler that attains fair- ness and high efficiency at the same time.
  • Approach
    • Fair time slice management, allowing I/O time slice fragmentation and concurrent request issuance.
    • Read/write interference management, blocking all writes until outstanding reads finish.
    • I/O parallelism
    • I/O anticipation
  • Evaluation
    • Fairness: some tasks run in parallel, each task experiences a factor of n slowdown compared to running-alone. We call this proportional slowdown. Average latency is used to measure the slowdown.
    • Efficiency: sum of (throughput(concurrent) / throughput(alone)).
• D T. Nguyen, et al. Reducing Smartphone Application Delay through Read/Write Isolation, MobiSys 2015
  • Contribution
    • A software solution that schedules read/write I/Os on mobile phone that prioritizes reads over writes to reduce application load time.
  • Approach
    • I/O priority assignment: make a new priority for delayed write, with 100ms maximum delay.
    • I/O grouping: group reads and writes, and dispatch reads before writes.
  • Evaluation
    • The system reduces launch delays by up to 37.8%, and run-time delays up to 29.6%.
• M Lee, et al. Improving Read Performance by Isolating Multiple Queues in NVMe SSDs, IMCOM 2017
  • Contribution
    • A software solution to isolate read and write requests in NVMe SSD.
  • Approach
    • Assumption: NVMe SSD controller processes requests in hardware queues in a round-robin manner.
    • Separate read and write requests into different hardware queues.
    • Moreover, group read/write hardware queues so that the rate of mix is even lower.
  • Evaluation
    • FIO read/write throughput (IOPS).

• M Nanavati, et al. Decibel: Isolation and Sharing in Disaggregated Rack-Scale Storage, NSDI 2017

• B S Kim, et al. Utilitarian Performance Isolation in Shared SSDs, HotStorage 2018

Open Channel SSD

• J Huang, et al. FlashBlox: Achieving Both Performance Isolation and Uniform Lifetime for Virtualized SSDs, FAST 2017
  • Contribution
    • A hardware-isolated virtual SSD solution achieving both physical performance isolation and wear leveling.
  • Approach
    • Use open-channel SSD to manage channels, dies and planes directly, thus achieve physical isolation except for some bus sharing.
    • Periodically swap most-wear and least-wear physically isolated units to achieve wear leveling.
  • Evaluation
    • Evaluating tail latencies.
    • Evaluating interference between latency sensitive and throughput sensitive applications.
    • Evaluating migration overhead.

New SSD Architecture

• J Kang, et al. The Multi-streamed Solid-State Drive, HotStorage 2014
  • Contribution
    • Make SSD accept multiple I/O streams.
  • Approach
    • An I/O stream accepts data with similar lifetime.
    • Data of different streams are stored into different piles of blocks. GCs are separated as well.
  • Evaluation
    • Evaluated the implementation on Cassandra, got higher throughput and lower tail latencies.
• J Kim, et al. Towards SLO Complying SSDs Through OPS Isolation, FAST 2015
  • Contribution
    • A new SSD architecture that separates OPS (Over Provisioning Space) among different tenants so that GC caused by different tenants are not interfered. This interference hurts because when data is mixed, a single GC might involve data from multiple tenants, and block them.
  • Approach
    • Assign different tags to data from different tenants
    • Inside SSD (implemented in SSD simulator), put data from different tenants to different regions
  • Evaluation
    • Proportioned multi-tenant I/O bandwidth test.
• S Yan, et al. Tiny-Tail Flash: Near-Perfect Elimination of Garbage Collection Tail Latencies in NAND SSDs, FAST 2017
  • Contribution
    • A new SSD architecture that nearly eliminates GC-blocked I/Os.
  • Approach
    • Plane-blocking GC: Finer grained GC that happens on a specific plane instead of die or channel.
    • Rotating GC: Orchestrating GCs in multiple planes so that they don’t happen at the same time.
    • GC-tolerant read: Use techniques like RAID5 so that data is available when one of the planes is blocked by GC.
    • GC-tolerant flush: Use hardware RAM based write buffer to store write contents temporarily.
  • Evaluation
    • Evaluating tail latencies.